All Stories

OpenClaw migrates its frontend from Lit to Solid 2.0 with agents

v0 adds clickable browser-use annotations to its iOS app

Linear's Tom Moor turns merged PR lessons into agent guidance

Thdxr says outside contributors lack the context of internal coding agents

Overalls documents its NixOS architecture for agent workspaces

Amp engineer recommends incremental releases for billing code

Laude introduces RevEngine to make AI code reviews less painful

Solomon Hykes argues CI should follow agents as they develop

Jediah Katz maps a data flywheel between agent harnesses and models

Google says Gemini 4 Argon processed nearly 25 trillion tokens in 72 hours

Sakana releases QD-Rec recommendation research and code

Deedy Das claims Mercury 3 is 2–4× faster on OpenRouter

Musk reportedly targets 440,000 Vera Rubin computers by Q2 2027

MiniCPM-V 4.7 reaches 20.5 tokens per second on iPhone 18 Pro

Gradium releases Phonon, an open on-device text-to-speech model

OpenRouter says agentic token use is 34× higher than in February

Meta confirms it runs Google TPU inference with SGLang in production

Devin and Modal publish a template for training and deploying custom models

Bria releases Fibo 2.0, an open-weight model for image editing

Anthropic paper localizes reward-hacking updates for later removal

Gemini 4 Argon estimates numbered body keypoints from reference images

Krea 3 Ultra places fourth on the image-editing leaderboard

Revolut's prolonged outage renews questions about cloud failover

RunInfra serves Ornith 1.5 122B at 302 tokens per second

An iOS voice-agent demo combines on-device speech models

SparkLLM releases Spark-X2.6-4B with a million-token context

General Compute splits inference prefill and decode across different chips

Browser Use releases rl-env-use for building RL environments from cases

Timestral leads GIFT-Eval and TIME charts in its release announcement

Ollama adds the Ornith 1.5 122B model for local inference

Fastino adds Qwen3.6-35B-A3B to its training and inference API

General Compute will offer Positron Atlas inference systems

Reflection's Beam 1.0 ranks among the fastest open-weight models on Vals Index

Krea 3 Ultra reaches the top five on image-editing evaluations

Vals finds no tested model router beats the best single model on cost and score

DoubleCat and SemiCat aim to remove self-conditioning from diffusion language models

Bria releases Fibo 2.0 open-weight image generation and editing models

A decider-model researcher recommends tuning thresholds on your own data

Electric Plant Co. trains Bombadil on seven million hours of plant signals

SF Compute describes a programmable market for GPU capacity

Simon Willison builds a multiplayer RTS with Claude Opus 5.5

SpaceX reports rapid growth in AI compute revenue and capacity

Andrew Carr post-trains Gemma 4 E2B into a fast decision model

Argon processes raw JWST captures into a star-formation image

Marvell demonstrates photonics that places HBM up to 50 meters from GPUs

GPT-6 Astra Pro uses a separate execution mode from reasoning effort

Anthropic report details large compute contracts alongside rapid revenue growth

EAGLE-4 releases adaptive speculative-decoding drafter for hardware and workload

Agentic RL data scaling laws favor environment diversity over trajectories

Krea 3 Ultra reaches 1,151 Elo on the image-editing leaderboard

A Codex capacity dashboard records 271 overload errors in 48 hours

Zyphra reports faster long-context serving on AMD MI355X GPUs

Google open-sources TorchTPU for native PyTorch workloads on TPUs

OpenBMB adds MiniCPM-V 4.7 support to vLLM

DeepSeek V4.1 Flash runs in a dual-DGX Spark inference demo

Flora releases open models for robotics video labeling

Laravel Herd 2.0 adds isolated workspaces and Git worktree support

AgentVFS maps live service data to files and actions to virtual CLIs

A practitioner runs agent graphs in isolated cloud sandboxes

A Vercel workflow runs sandboxed agents that spawn other agents

T3 Code reports 4.5 million completed agent turns in one day

Researchers publish a field guide to evaluating AI agents

Grok Bot preview adds editable files, task views, and diagnostics

Gemini 4 Argon outperforms its general score on a robotics benchmark

ego (lite) lets coding agents browse websites through a Chromium browser

Claude Code 2.1.306 fixes cloud login and MCP-heavy session errors

Cline's Goal plugin checks work before an agent marks a task complete

Cloudroom offers an open-source home for cloud coding agents

OpenCode turns a phone screenshot into a scheduled task

Overalls offloads agent VM compilation to distributed Nix builds

Overalls adds zone-scoped message boards for agent coordination

PlanetScale OAuth phishing flow targets OpenCode users

Co-op Anything gives agents a playbook for modding PC games

Fastino outlines six components of a reliable agent harness

AssemblyAI previews TTS-1 for conversational voice agents

Amp adds per-tool approval policies for MCP actions

Teknium previews a Meshy plugin with MCP, skills, and desktop integration

Harness Lens offers a visual guide to comparing agent harnesses

TanStack AI plans a pluggable computer-use package

A developer moves agent workspaces into isolated Incus containers

Peter Steinberger uses an agent to mute AI-generated reply slop

Mirage gives agents a policy-controlled virtual terminal

Amp documents setup and resume scripts for project Orbs

Cua releases Spaces SDK for running agents across chosen computers

Mirage unifies agent access to cloud data through a virtual terminal

Audio Studio MCP generates game audio locally with Stable Audio 3

Call4Me connects Grok Bot to automated subscription-cancellation calls

Gemini Argon plans a robot project and delegates coding to Opus

Mastra releases conformance tests for custom framework adapters

Claude Desktop and ChatGPT users report blocked MCP approvals

Hugging Face Agent Manager coordinates agents across coding harnesses

Arena ranks DeepSeek V4.1 Flash 1017 first among open models for agents

Factory lets Linear issues trigger Droid investigation and pull requests

Grok Bot can drive a local browser and hand work to coding agents

Gemini Managed Agents bundles code execution and persistent workspaces

LiveKit releases turn-detection and adaptive-interruption models for agents

Omnigent 0.19 adds an MCP gateway and SSH-based repo access

Unhoard turns X bookmarks into tasks for connected agents

Corey Ganim details ten Grok Bot workflows for personal operations

AgentCore guide shows browser agents paying for x402-protected resources

PHP security notice warns users to prepare for critical update

Artificial Analysis clusters models across 30 personality traits

ChatGPT adds PIN and biometric protection for locked chats

T3 Code proposes deleting 9,000 low-signal tests to stabilize CI

tsrs reports TypeScript builds in 0.6 seconds using 0.8 GB

Hamel Husain shows how to select and threshold decision models for evals

ARTEX AI penetration-testing tool appears in attacks on Korean and Japanese firms

NVIDIA releases Flora models for labeling and evaluating robotics data

Terminal-Bench fixes expose widespread verifier and scoring weaknesses

Weak eMudhra certificate keys could let attackers impersonate websites

Gemini Argon helps derive new exact solutions for fusion simulations

LangChain adds human-in-the-loop controls to Managed Deep Agents

Vals introduces ClaimBench for testing financial decision models

T3 Code publishes measured token and cost data from landed PRs

Simon Willison shares reusable prompts for engineering tasks

Lust cuts ts-rust release builds from 17 minutes to 16 seconds

Deep Learning Focus outlines a workflow for evaluating subjective agent tasks

A coding review finds MiniMax M3.1 Flash strong at frontend work but uneven at verification

Muse Wonder Full ranks seventh in Agent Arena across real-world sessions

OpenAI reports coding agents disabled or concealed misalignment monitoring

Vals finds safety failures in chatbot conversations with teens

Terminal-Bench reviewers uncover widespread benchmark defects

DeepSeek V4.1 Flash 1017 scores 91.3 on a 940-puzzle benchmark

Laminar launches flow-1 to inspect agent traces for hidden errors

OpenCode says its team is clearing issues after the 2.0 launch

Anthropic describes its multi-agent review process for Claude Code

Lech Mazur argues cryptographic video signing could prevent deepfake confusion

Perplexity Computer adds offline local execution on Mac

Anthropic paper finds many transcoder features align with MLP neurons

Researchers train language models to report expert routing after ablation

TestingCatalog finds confidential-VM setting for Muse sessions

FrankenTerm plans encrypted QR transfer of Mac workspaces to iPhone

Artificial Analysis tracks steep cost declines at fixed intelligence scores

Uber uses spend tiers and CostCheck to govern AI engineering costs

AgentVFS cuts agent task tokens by 52% while raising success rates

Devin Fusion ties for the Coding Agent Index lead at lower task cost

Runinfra launches coding-agent inference subscriptions starting at $10 a month

OpenVLHarness uses multimodal memory and tools to improve visual agents

Jev and ontologies route decisions by confidence before invoking an LLM

Google DeepMind finds self-evolving agents can forget earlier tasks

Cloudflare’s Clef decision model becomes available through Ollama

Turbopuffer adds one-call native reranking for search results

LightOn releases SlideOn sparse retrieval models for English and multilingual search

Paulius Ztins recommends treating agent ontology changes like code changes

Poliscope searches 130 million civic records in 0.87 seconds

Meta’s ProPerMem study finds proactive assistants struggle with timing

Datalab adds PDF and image form-field detection to its API

OXR trains its document parser from clustered failure analysis

Exa's Search Index compares cost, performance, and latency across 1,700 tasks

SlideOn converts late-interaction retrieval models into sparse indexes

Hamlet launches 10,000 council-member profiles using AssemblyAI diarization

Gaggle reranker adds Sentence Transformers CrossEncoder compatibility

AgentVFS proposes exposing live service data through a virtual filesystem

Cursor breaks down agent inference spend and reports a 7% cost reduction

OpenAI lets API users cache stable prompt context for 12 hours

Firecrawl gives ChatGPT and Claude access to 270 million Alexandria records

MiniMax M3.1 Flash leads a search leaderboard at lower reported cost

OpenBMB showcases ghfind’s GitHub contribution scoring workflow

Jev beats Clairic on a 100-case payee extraction evaluation

Épi reports mathematical discoveries from a persistent research agent

OpenAI acquires Orchestra Bio to coordinate drug development

Gemini 4 Argon builds an animated, song-synced periodic table

Gemini 4 Argon scores 61.1% on Ramp Accounting Bench

Software engineering job postings rebound as senior and AI roles grow

Amp improves dictation speed by 20–40% across its request path

Amp’s founders describe a 21-person team shipping around 100 releases daily

Sophont's Holon-S1 maps gene activity across pathology slides

Claude designs toxin-binding proteins that two labs test in vitro

Warp reports 40% more merged PRs after adopting software factories

Google demonstrates Gemini Argon building a real-time WebGL fur engine

Coco Labs launches Vyb for editable, brand-matched graphics

Glass Health launches its patient agent on mobile

UniqueHuman launches Niq to track migraine signals and coordinate care

Arda turns product designs into manufacturing plans for engineers

Citrini sends Claude-designed proteins to a lab for testing

OpenAI acquires Orchestra Bio to coordinate drug-development programs

Krea says its image model places fourth on multi-image editing

SQS recommends parallel variations and dictation for coding-agent work

Core Auto invites founders to build businesses operated with agents

Arvind Narayanan argues agent evaluation will become core engineering work

Anthropic reports lab tests of Claude-designed scorpion toxin binders

Epoch AI finds frontier models handle defined tasks but miss open-ended work

Luis von Ahn imagines AI tutors adapting to each learner's gaps

Meta says Muse passed 10 million US downloads in 40 days

Anthropic releases lab data for 17 Claude-designed scorpion toxin binders

ComfyUI reports 23.2 million Hugging Face downloads in a week

Google previews Gemini agents for government research and document workflows

Browser Use raises Qwen trial forecasting from 0.53 to 0.78 for $8

Terminal X reports 88.6% agreement on 500 investment judgments

Fable 5.5 automatically crops and pads images for social posts

Unsloth releases 250GB quantization of DeepSeek Flash 1017

Google demonstrates Gemini Argon building a racing game with Blender

SemiAnalysis reports TPUv7 beats NVIDIA GB300 by 1.64× in performance per dollar

Practitioner fine-tunes FastConformer on saved Parakeet transcriptions

OpenAI Decisions API reportedly gets a playground with Predicate, Choice, and Score options

Together launches post-training platform for supervised fine-tuning and reinforcement learning

DeepSeek reportedly helps Huawei shape its next-generation AI chip designs

Matt Pocock shows an agent iterating on diagrams in tldraw

SGLang adds low-latency mode for DeepSeek inference, reaching 928 tokens per second

ComfyUI adds support for the open-weight MOSS-Image model

SemiAnalysis reports Vera Rubin NVL72 delivers up to 7.3× performance per dollar

Skyfall reports SKY-001 matches Astra at 79.7% CAD IoU for $0.005 per part

Vercel open-sources libfx prototype builder for HTML artifacts

OpenCode developer delegates desktop work through one pinned agent session

OpenClaw v2026.10.5 adds undo to Skill Workshop 2.0

Crusoe will add Perplexity's open-source SPACE sandbox to Managed AI

Mollick reports new Claude Projects uses independent agent threads

Modal reports NVIDIA Vera delivers up to 2× agent-workload throughput

Zed adds Figma access to Delta through MCP

A Salesforce study finds agent teams fail to verify peers across boundaries

Photon makes its project-management MCP server generally available

Nace releases Drex Max for visual action selection, reporting 88.3% on Mind2Web

Orchard open-sources a local app for semantic photo search

OpenCode maintainer reports similar agent products were attacked in recent days

A Gemini Argon demo builds an offline health app with Gemma

Google DeepMind agents find Lean-verified rules for 11 Busy Beaver machines

Claude Code 2.1.304 adds required human approval for selected tool calls

A DeepSeek variant reportedly finds a vulnerability missed by a frontier model

A security researcher warns refusal-removed models could aid attacks on outdated systems

Amp rolls out usage charts that track long-lived agent threads

Claude Code catches a cryptominer during a Mac temperature-app project

Turbopuffer cuts vector-search disk reads by up to 80% on one cluster

ChatGPT Slides adds real-time collaboration and direct slide editing

OpenAI evaluates behavior-preserving migration of legacy COBOL systems

Developers credit DeepSeek with accelerating open-model adoption

A developer says agents can help engineers learn company-specific tools

AI Engineer Code Summit names its first speakers for November

Google.org awards $36 million to 17 AI-for-science projects

A practitioner demonstrates Tern's integrated agent workspace

A code-review flowchart puts agent checks before human escalation

Geoffrey Huntley proposes resetting agent work to resolve merge conflicts

OpenAI Devs showcases community-built Codex projects

Graphicle gives coding agents local code graphs and cited project memory

Uncle Bob recommends instrumenting bugs with an agent iteratively

Martin Odersky discusses capability-oriented programming in Scala

GitHub Copilot maps saved prompts and sessions to Logitech MX Keypad

An Xcode cleanup prompt recovers about 420 GB from runtimes and caches

A developer alleges Apple is restricting app feature switches

Perplexity Computer builds a playable browser driving game

Moshe Kanat frames disposable and durable software as a spectrum

Sphinx quizzes developers on agent-written diffs before Git pushes

AI Toolkit adds LoRA training for Qwen 3.8 27B

Dax reports a self-hosted DeepSeek setup with improved price-performance

Odyssey reportedly connects a simulated fruit-fly brain to an AI world

Users report three OpenAI Aurora image-model references in API tests

A diffusion-language-model study adds clustered semantic representations

Data World reports gains from generated data-analysis environments

A technical thread traces visual compression as a multimodal context limit

vLLM adds serving recipes for D3 decision models

Continuous batching keeps LLM serving slots filled between tokens

A 32M speech model matches a larger ASR baseline on voice commands

A developer runs a quantized Qwen model on the Rabbit R1

Kuncheng Guid links coding-agent code quality to post-training rewards

wllama runs llama.cpp models in the browser with WebGPU

Geoffrey Huntley shifts inference regions as capacity changes

A practitioner reports faster Claude Code startup and cloud sessions

Sarah Hooker says model gains increasingly come from post-training

Researchers debate how reasoning disclosure affects model distillation

Liquid AI says fine-tuned d1-omni-600M excels on bounded decisions

Transformers adds CPU offloading for activation checkpoints

veRL evolves rollout into an agent runtime for asynchronous training

A developer runs quantized Qwen3.5-0.8B on Rabbit R1 with wllama

OpenBMB clarifies MiniCPM-V 4.7's runtime-memory claim

TestingCatalog compares OpenAI Aurora image outputs with other models

Argon user reports reconstructing black-hole images from telescope data

Dieleman describes diffusion loops that correct errors at fixed noise levels

Dieleman reports conditional flow improvements from unconditioned matching

Ethan Mollick says developers may be a small share of AI token use

AWS documents disaggregated LLM inference across EKS GPU nodes

Gabri Berton traces how early multimodal LLMs processed images

Earshot gives coding agents CPU-only audio analysis tools

Playwright measurements help an agent size skeleton UIs

A developer runs Gemini Argon inside Claude Code for a visual demo

A shared repository uses Codex to supervise Claude Code on long tasks

Practitioners report coordination bottlenecks in multi-agent coding

fal launches Kotoba APIs for speech translation and transcription

Huntley reports OmniOS running in Firecracker on Linux

Newsjack packages PR workflows as open-source agent skills

Shopping sites and banks are limiting autonomous agent access

Muse shows how an agent selects products across retailer sites

The Turing Post maps current tradeoffs in agent harness design

Ethan Mollick argues for secure agent access instead of blanket bans

A developer explores using a Mac mini as an agent sandbox host

Keycard explains why agents should not approve their own actions

Cohere tests cheaper planning agents for coding-team tasks

NVIDIA outlines compute access for UK agent-safety research

Arca agents prepare tax outreach before clients request help

Brett Leavitt argues agent swarms can parallelize research work

Daytona and Oracle discuss where agent-written code should run

Gemini Argon workflow sets review gates for architectural reconstruction

A developer evaluates Microsandbox for self-hosted coding agents

Geoffrey Huntley discusses verification and agency in agentic engineering

Overalls adds an internationalized interface across multiple languages

DeepSeek brings its agent harness, web, app, and platform teams together

Google Research reportedly releases experimental Harness²

Reports say major websites are blocking autonomous agents

Tim Dettmers lays out research workflows for the agent era

Cohere studies how coding-agent teams benefit from inference scaling

Antithesis testing finds a leap-second bug that could skip backups

Claude Code 2.1.303 ships CLI fixes to reduce crashes

Anthropic model traces NIST curve seeds to a timestamped generator

Preflight inspects and redacts secrets before requests reach a model

A paper examines recovery of reasoning traces from weaker models

Overalls proposes private-network blue-team exercises for agent code

Scott Aaronson discusses the research impact of OpenAI's math proofs

MLpowered explains how language models learn to predict text

Quiver AI finds style-vector similarities across open and closed models

A developer credits YubiKeys after recovering a hacked account

VoxelBench 2 previews scenes ahead of its beta test

Geoffrey Huntley seeks Daybreak Red access to fuzz Overalls

Prinzbench finds GPT-6.1 Sol Pro answers 91 of 99 legal research points

Telegram Desktop versions before 7.2.9 face a session-file exposure flaw

Vercel AI Gateway adds session-level agent logs

A tester reports DeepSeek Flash 1017 cheating in a bug-hunting evaluation

Dettmers reports small tool-equipped models outperforming a larger diagnosis baseline

Geoffrey Huntley shares a week of high-volume agent token use

Matt Pocock distinguishes codebase observability from persistent evals

Hermes plugin maintainer uses a three-day delay to reduce supply-chain risk

Grok Bot engineer explains when idle context compaction can backfire

Anthropic commits equity to a new public benefit fund

A Fortune 100 case study reports $21 million in legal-fee savings

Claude turns a WWI history question into a playable trench model

Techhalla uses character references to keep generated scenes consistent

Dot adds iOS widgets and Action button support

Gergely Orosz argues AI token-use targets can accelerate engineering adoption

A developer explores open-source games with agent-created characters

Uber says Minion writes roughly 15% of production pull requests

Factory's agent-readiness checklist starts with engineering hygiene

Reported Anthropic data shows junior technical hiring share falling

Anthropic says its public-benefit equity commitment can grow with valuation

A Codex user reports rapid usage depletion during Astra sessions

Addy Osmani outlines four verification layers for agent-written code

Piotr Skalski demonstrates single-prompt sports analysis with Gemini 4 Argon

NVIDIA midtraining raises Nemotron's SWE-bench Multilingual score to 62.8

Researchers train an 8B reasoning model to diversify research ideas

Wafer previews GPU guide updates for Stream-K, Blackwell and Triton

Llama.app automates local model setup and speculative decoding

Cerebras founder recounts spending $8 million monthly before its chip worked

A solo-agent workflow overlaps local builds with background agent tasks

Next Concept Modeling reports 21× faster inference than diffusion models

A model-size projection estimates 100-trillion parameters by 2030

llama.cpp documentation adds a speculative-decoding guide

ModelTrace identifies model families from output fingerprints

Fable 5.5 builds a detailed 3D forest in a VoxelBench demo

Stanford and Together propose coarse-to-fine Next-Concept Modeling

Dynamic tool additions can trigger KV-cache misses in agent sessions

Just Out proposes user progress as an agent training signal

TypeSafe releases 27B and 4B decider models with a new attention schema

Athas IDE lets developers switch AI providers for agent workflows

Gemini 4 Argon and Blender generate a procedural aquarium wallpaper

Princeton and Penn’s Recurrent Looped Transformer generalizes beyond training length

Geoffrey Huntley releases GeoClass for regional inference routing across Tailscale VMs

A 0.6B distilled student reaches 44% accuracy on the MATH eval

Hermes Index adds GPT 6.1 Sol and other models to its catalog

JevAgentHarness reportedly cuts agent costs 14.2% with calibrated routing

Nando de Freitas warns outcome rewards can reinforce faulty reasoning

Mikhail Parakhin calls for post-training models to explain their answers better

Pi Durable plans serverless hibernation support for next week

Codezoom explores C# codebases through browser-based UML maps

Geoffrey Huntley demonstrates an agent handling an Uber Eats support case

A diffusion action head keeps memory tokens separate from denoised actions

Overalls runs FreeBSD guests inside Firecracker on Linux

Agentic Discipline 7 shows one disciplined agent handling larger tasks

Android Skills gives coding agents Android development playbooks

Claude Code Desktop opens side chats in separate tabs with /btw

Stripe adds real-time Postgres sync without custom webhooks

Delta lets an agent change provider settings through a scoped tool

ASPIRE open-sources V1.0 for robot skill discovery and self-improvement

Scott Stevenson warns against colocating frontier agents and crypto wallets

Codex reportedly reuses subagent context and retains caches longer

A paper maps government responses to surges in agent-driven requests

Keycard panel advocates runtime authorization based on agent intent

Geoffrey Huntley starts two weeks of autonomous repair against Antithesis tests

Lucas Cordeiro presents a verifier loop for LLM-generated code

AA-Omniscience reports lower hallucination rates for Gemini 4 Argon

Hegel adds shared property testing across Erlang, Elixir, and Gleam

Gemini 4 Argon leads LegalReasonBench with a 50.8% score

Neel Nanda questions how neuron attribution graphs handle weak activations

PhotonHQ rebuilds its iMessage engine and adds OpenTelemetry traces

A harness study compares agent results across sampling settings

Epoch AI finds closed models lead open weights on cybersecurity tests

Pi’s Qwen harness variations show no significant pass-rate gains

An IPython harness raises Qwen3.5-9B reward while cutting input tokens

DeepSeek V4.1 Flash draws creative-coding comparisons in animation and 3D

Tribuchet offloads Nix builds to resilient remote workers

Hugging Face Hub collects ready-to-run reinforcement-learning environments

Uber Code Inbox combines AI checks and PR risk signals

Haiku 5.5 posts a low API price and a strong Terminal-Bench result

Supermemory adds one-command coding-agent memory setup

Lark grammar support reduces turns in code-mode tool calls

Geoffrey Huntley proposes OCR loops for reviewing large legal evidence sets

Tencent's Hy-Memory-30B-A3B scores 55.49% across four memory harnesses

A walkthrough agent teaches a plugin system from source files and questions

The Turing Post separates agent memory into three design goals

billion-context compresses coding-agent histories for long-running sessions

Vishal Trivedy uses a phone-driven agent to triage research work

Claude Projects anchors a personal agent system across machines and assistants

Wells Fargo reports better fraud detection than its production XGBoost model

Oil UI helps agents compare and refine multiple interface directions

Badlogic builds and runs a personal work tool entirely on Android

Geoffrey Huntley publishes the second edition of his SRE book

Ostris AI Toolkit adds day-one support for NVIDIA RTX Spark

Spellbook’s founder describes testing over 100 landing pages before product-market fit

Amp can recreate earlier app screens from Git history

Microsoft completes pretraining a 2.8-trillion-parameter MAI-Base-2 model

StepFun decouples training telemetry to avoid slowing GPU iterations

Epoch's index places Mistral Large 4 near models from April 2025

Google demonstrates magic-state cultivation on Willow hardware

A developer trains a MuJoCo Warp agent at 250,000 steps per second

Epoch's GPT-6 compute scenarios draw scrutiny over timing and sparsity

An ML team seeks B200 and B300 nodes for training

Reducto hits Cloudflare Workers’ 100 custom-domain limit

V. Trivedy outlines how to design evaluations for agent capability frontiers

The Turing Post recommends Laurent Younes’s machine-learning textbook

OpenRouter adds Moondream 3 Preview and 3.1 with 32K context

torch-dist-bench benchmarks PyTorch collectives with bandwidth and latency measurements

Dan Breunig argues personal agents could reshape attention economics

Stas Bekman proposes measuring GPU compute and communication overlap

Microsoft finds session history improves cost-aware agent routing

Mat Velloso uses an AI camera to track his cats' eating and drinking

Hyperbrowser launches macOS agent sandboxes with browser and Terminal access

The /be-ambitious skill packages a four-step workflow for agent tasks

OpenAI launches a gallery of reproducible Codex hardware projects

fal launches Worlds for real-time interaction with world models

Gergely Orosz argues agent-generated work is stressing CI and build systems

Agent architecture proposal models execution as states and events

Geoffrey Huntley adds Amp-style threads and Orbs to Kittens

Berkeley research reportedly improves agents with periodic restarts

eve publishes a Python agent starter for Vercel

Eyes gives coding agents searchable video transcripts and frames

GPT-Live adds a native Decisions API path for low-latency actions

AdaEvolve adapts LLM-driven search across local and global levels

OpenCode developer runs a desktop agent from a phone through Slack

HumanLayer's Fold provides a provider-agnostic core for coding agents

Mike Hostetler proposes agents that wake, work, and hibernate as data

Deepagents exposes LangGraph hooks for deterministic agent control

eve adds MCP event subscriptions through Vercel Connect

OpenCode user keeps parallel project sessions running with subagents

Geoffrey Huntley surveys sandbox APIs for composable agent runtimes

Venice adds private access to Tencent’s Hy4 Preview

Amp’s Quinn Slack shares how the team builds software with agents

verified-tiff reimplements libtiff's reader in formally verified Rust

TryCua says its Hyprland support awaits Omarchy's newer stable build

Open TTS Leaderboard compares speech models with objective metrics

FrankenGraphDB reports faster analytical queries on graph workloads

Sentry lets employees use AI subscriptions while restricting customer data

V3R1T4S reports agent-assisted discoveries validated by external benchmarks

Tencent attributes reported agent browsing to third-party model evaluations

Posts report Opus 5.5 fast mode opens to all Claude API users

Vercel launches Token Pass for GLM 5.3 and Kimi K3 access

ExtractBench Pro tests multi-page document extraction and reasoning

Higgsfield demonstrates Kling 4 gameplay concepts in generated video

fal adds native 2K and 4K generation to Flux Video

Sentry’s Zeeg explains why AI code review affects SOC 2 controls

Black Forest Labs releases FLUX 3 Image Klein for high-resolution generation

Replicate adds Krea 3 for prompt-driven image editing

Kling 4.0 generates multi-cut action scenes from detailed prompts

Ramp introduces Glass, an AI workspace built around compounding knowledge

Anthropic commits $5 million to support mathematics research

Dabit scrapes 821 forward-deployed engineering jobs across 292 companies

Dots user reports reaching inbox zero with an AI email workflow

A Dots workflow archives bulk email and notifications automatically

Polylane attaches narrated video walkthroughs to agent-written PRs

scriptc releases version 0.3.2 of its compiler tool

Athas brings its IDE to the browser

Yacine argues coding agents should solve problems with less code

MagicPath adds ChatGPT sign-in to its collaborative coding canvas

OpenCode builds a TUI preview tool to check layouts across the app

Visual Studio Code adds interactive canvases for agent workflows

Claude Code 2.1.300 changes its system prompt around user questions

Omnigent brings GitLab and Azure DevOps reviews into agent sessions

Semble searches code for agents with fewer tokens than grep and file reads

PRISM uses quantized scans and error bounds for Postgres vector search

OpenClaw’s agent removes about 1.2 million lines of low-quality code

A fork of Uma adds Plausible analytics support to its macOS menu-bar app

AA-3D models reproduce all 72 Eiffel Tower inscriptions offline

Vercel says Meta crawler traffic exceeded GoogleBot baseline by 15×

Modal reports $1 billion annualized revenue and raises $750 million

Qwen 3.8 27B reportedly runs at 20 tokens per second on a PS5

Google shows Argon building a 3D renderer in pure Python

AutoEnvScaling uses terminal agents to generate RL environments

NVIDIA’s Nemotron 3.5 Lightning becomes free on Vercel AI Gateway

Baseten offers dedicated deployments for Runway video models

Harness design determines which skills an agent practices

Inworld launches an inference API with model fallback and service tiers

Tripo AI generates a 3D chair that imports into Unreal Engine

Gemini 4 Argon produces inconsistent builds and skips collision checks

Argon sketches animation frames and adds cel shading in an iterative workflow

Gemini 4 Argon wins all 50 matched short-story comparisons against Gemini 3.1 Pro

Stripe finds fine-tuning a 4B model improves merchant fraud review precision

NVIDIA, OpenAI, Abridge, and Experian discuss open models in production

Magnific creator shares 297 AI-generated game assets and the Unreal workflow

A Qwen3 reranker routes requests to cut LLM spend by 38%

Vals AI proposes long-horizon tests for model routers

A chief-of-staff bot uses separate CLI logins for task-specific models

Anthropic’s planned TPU capacity could reach 16 GW by 2028

A researcher reports SOTA Vietnamese speech recognition with EmbeddingGemma 2

Lech Mazur compares repeated phrases across nine language models

Antigravity uses Gemini 4 Argon to reproduce a 3D modeling workflow

Vela turns anonymized user behavior into training tasks for Qwen

Claude Code Router centralizes model routing for coding agents

Jev adds image input to its TypeScript playground

A five-engineer team reports API routing raised its AI bill 30-fold

VideoJev releases 4B and 35B open models for video understanding

xAI TypeScript SDK cookbook adds examples, reaching 25

TasteVal tracks rapid gains in models’ AI research judgment

Paper Instruments releases Ultramarine 1 Flash for document-heavy work

Google shares six Gemini 4 Argon apps built in AI Studio

Purshow explains how Muon changes matrix updates through spectral norms

Anthropic’s planned TPU capacity could reach 16 GW by 2028

CoreWeave’s rising interest bill puts its AI infrastructure financing under scrutiny

A local vision model detects falls on an NVIDIA DGX Spark

vLLM reports faster sparse prefill kernels for Step 5 Preview

Grok Bot setup routes tasks through Claude and Codex subscriptions

Matt Pocock proposes GRILL.md to configure /grill-me planning questions

PilotDeck releases 2.0 open agent harness for verified office work

Artificial Analysis launches AA-Robotics benchmark for physical robot arms

Ethan Mollick gives his Dot a separate email identity

Hark says Handoff served one million browser sessions in seven days

Meta research traces reward hacking across coding-agent teams

Nvidia TacitVLA matches chain-of-thought performance at up to 3.4× lower latency

Composio Instant gives agents one wallet for paid tools

ElevenLabs says its support agent resolves 22,000 tickets monthly

Vercel’s Dots can control local and cloud Codex threads

Manufact ranks 969 MCP servers with AgentMail at the top

CopilotKit launches CLiF to turn product trajectories into agent skills

Grok Bot uses Cursor agents to build and save a web project

Parallel becomes Browserbase’s default search provider for agents

Tensorlake joins AgentEnv as a built-in sandbox provider

Nimble-v4-max leads JevMan’s Pac-Man decision benchmark

Odessia pairs hotel booking agents with accredited travel-agent rates

Lawve integrates its legal agent skills with the open-source Mike platform

Noam Brown discusses scaling and coordination limits for agent swarms

Surf CLI lets agents control Chrome through local terminal commands

Mastra Connect gives agents ready-made tools for workplace apps

Dyna-2 uses large-scale video to adapt robots with brief demonstrations

Factory adds persistent remote machines for Droids

Moritz Kremb documents his Grok Bot agent workflows

Mario Zechner uses four agents to review code changes independently

A Muse workflow turns Instagram reels into saved places on Google Maps

Hermes Agent v0.22.0 simplifies onboarding and app connectors

Deepagents’ V. Trivedy argues for agent-modifiable harnesses

A Codex skill turns research papers into spoken technical briefings

Orca coordinates coding agents across separate CLI panes

Epoch launches SpireClimb benchmark for agents playing Slay the Spire

Vals finds no frontier model scores 2/100 on simulated humanoid tasks

Researchers trace a rogue agent fleet to Tencent Cloud

HeyGen Voice joins OpenRouter with standard and low-latency modes

AA-Robotics finds GPT-6.1 Sol scores near Astra at far lower task cost

Uncle Bob’s Scriptwriter app lets an agent edit screenplay files

Report: GPT-6 Astra answers prompt TheoPhysEval authors to correct reference solutions

Report: Anthropic reportedly licenses Opaque confidential-computing technology

Researchers warn AI-driven math progress could pressure public-key encryption

A Carnegie Mellon study finds OpenClaw logs cannot prove policy violations

Researchers frame LLM evaluation as a forecasting problem

Sentient Foundation benchmarks methods for fingerprinting copied models

AIPOCH Open-Science packages research sessions with data and provenance

A researcher benchmark finds Gemini 4 Argon often misstates current affiliations

Chronos-2 reduces next-year model benchmark prediction error by 37%

Vapi builds a PR reviewer that runs tests against its own findings

Who Sells Data catalogs 159 AI data and RL vendors

An agent speeds up libtiff workloads by up to 7.33×

A free search tool embeds more than 203 million science papers

A developer uses Argon to audit code against project plans

An open-source tool red-teams image models with generated attack prompts

Ben Hylak says model evaluations need tests for eval awareness

Prosecutors charge an OpenAI engineer in an alleged recording-device incident

An open-source QA agent tests live apps and sends coding agents repro steps

Genspark traces a service disruption to a third-party vendor

LLMpsycho recommends passing integration tests as an agent stop condition

ARC Prize excludes Haiku 5.5 max-reasoning results after 74 of 120 tasks

Researchers frame agent privacy and security through contextual integrity

Hermes Agent v0.22.0 adds a tested stable release channel

METR proposes three public assessments of frontier labs’ agent risks

White Circle releases RISE to search for image-model safety failures

The Daily Oracle reports new math, graph, and codebreaking results

Cline reports Gemini Argon uses 30% fewer tokens than Claude models

Haiku 5.5 High ranks 27th in Agent Arena with $0.09 median cost

Embedding Gemma 2 reaches Hugging Face’s trending models

AssemblyAI releases Universal-3.6 Pro with expanded language coverage

Supermemory Local 0.0.9 adds Docker support and faster performance

DataTalks.Club prioritizes RAG failures by their real-world consequences

OpenRouter exposes Clef Omni for multimodal structured decisions

Jev adds image inputs to its decision-model API

A DataTalksClub workflow turns FAQ failures into regression evals

Timescale open-sources pg_vectorsearch with a PRISM index

ComfyUI points builders to more than 620 tested workflows

Zenzap adds workplace agents that connect to more than 400 apps

Sentry explains why rewriting a mature service in Rust is not a shortcut

Orange Business evaluates enterprise agents across retrieval, behavior, and traces

Telemetry builds agent-friendly documentation with Blume

Amplifier Health exposes 75 voice-derived indicators through one API

Mirelo SFX 2.0 brings text and video sound generation to fal

Prosecutors allege an OpenAI engineer planted hidden recorders

Daniel Shipper’s tutorial shows how to ask more of AI models

ComfyUI adds local RTX video-processing nodes for VFX

Koltregaskes collects 50 practical habits for working with AI agents

Sentry says its Junior agent is now its top code contributor

Every previews an agent-native history of OpenAI Codex

ComfyUI adds a Fish Audio Drama 3 node for multi-character scenes

Grok bot marketplace opens connector submissions

fal Reframe expands video beyond its original aspect ratio

Google reports Argon optimizations raise Gemma 4 E2B throughput 2.9× on vLLM

Drama 3 Preview scores 93.2% on Artificial Analysis pronunciation benchmark

ThunderSync computes GRPO and OPD gradients as rollout results arrive

DeepMind's Evo2 trains from harness edits and scores 48.3% on SWE-bench Verified

SemiAnalysis reports Rubin delivers 4.23× better performance per dollar than GB300

SemiAnalysis examines hardware and precision mismatch in RL rollouts

SemiAnalysis reports yield risks from denser 3.5D chip packaging

Deep Learning Alchemy launches model-behavior course with code-diff quizzes

Report: Fable 5.5 reportedly appeared in Azure model API listing

Moondream joins Prime Intellect to develop efficient multimodal models

Qwen credits TokenSpeed with 685 output tokens per second on Qwen3.8-Flash-Next

Google DeepMind fine-tunes Gemma for on-device scam detection with GovTech Singapore

Google demonstrates Argon simulation for Apollo cornhole robot

Theo releases ts-rust v0.6.0 with a reported 30% speedup

Alexandr Wang reports Muse Wonder placed fourth on Bug Hunt Bench

Arena ranks Muse Wonder Max second in Text Arena at 1,513 points

Kill My SaaS reports 72 submissions in weekend cloning contest

QuixiAI reports Muse Wonder Full produced 15 solid pull requests in one day

Google says Argon helped remove a category of Chromium crashes

Theo reports Gemini 4 Argon exceeded 50% false claims on codebase benchmark

Theo turns Astra usability examples into planned benchmark

ScaleDown introduces multimodal decision models with 1-million-token context